{"id":"bc2f1e1a-3969-4cea-b6f8-bda0fc2028a3","arxiv_id":"2506.14865","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"ALSPG combines spectral projected gradient descent, an augmented Lagrangian, and geometric projections to solve constrained robot motion planning problems faster than common second-order solvers.","lead":"ALSPG is a first-order optimization solver that uses geometric projections, like moving a point to the nearest spot inside a box or circle, to handle constraints in robot motion planning. The paper reports faster solving than second-order methods such as iLQR on both constrained and unconstrained tasks, tested in simulation and on real robot arms and a small car.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Robust IK experiment reports an inverted chance-constraint violation rate (≈80% violation at η=0.8), suggesting ALSPG may fail to enforce a simple SOC constraint; this bears on the solver's validity.","rationale":"I considered whether the missing convex convergence theory alone is the main risk. It is real but explicitly acknowledged and common for practical first-order MPC methods; empirically, ALSPG works on many tested tasks. The more damaging issue is the Robust IK result, which is internally inconsistent and directly contradicts the mathematical meaning of the chance constraint. This is not a disagreement with external consensus; it is a checkable correctness failure inside the paper. If the 80% figure is real, the projection-based treatment of SOC constraints is wrong. If it is a typo, the paper should correct it and re-run. The reader's conditional verdict remains appropriate: the speed claim depends on the constraint-satisfaction claim, and this experiment casts doubt on it. I am not accusing the authors of dishonesty; the likely explanations are a typo or a sign error, both of which are addressable by a concrete check.","tokens_in":11761,"tokens_out":7375,"duration_ms":76317,"concrete_test":"Rerun the Robust IK experiment with η=0.8 using the released code: sample the line slope 1000 times from N(µ,Σ), compute the rate of a^T f(q) > 0, and check the sign convention in the SOC inequality and the Π_C projection implementation. If the violation rate is ≈20%, the '80%' is a typo and the concern is resolved. If it is ≈80%, ALSPG is enforcing the wrong probability level, and the speed comparisons for this task class are not meaningful.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the central claim that ALSPG handles constrained problems faster, the algorithm must actually satisfy the constraints in the reported benchmarks. The paper's own limitation statement (Section IV-B) concedes that the convergence theory from [12] is only for convex functions and convex sets, and that the general nonlinear/nonconvex extension is empirical. A concrete red flag is the Robust IK experiment (Section V-A): with η=0.8, a correctly enforced chance constraint a^T f(q) ≤ 0 should be violated about 20% of the time, yet the paper reports 'a constraint violation percentage of around 80%, as expected.' This is the exact inverse. Either the SOC projection/update in ALSPG is not enforcing the intended chance constraint (so the method returns solutions with the wrong safety level), or the evaluation is mislabeled. In either case, this experiment does not validate that ALSPG handles this projection-based constraint class correctly, which is load-bearing for the claim that projection-based constraints improve performance.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ALSPG, a first-order augmented Lagrangian method with spectral projected gradient descent and geometric projections, for constrained robot motion planning. It formulates constraints as Euclidean, polytopic, or implicit projections, solves optimal control problems via direct shooting with an efficient recursive transpose-Jacobi product, and validates the method on inverse kinematics, pusher-slider MPC, obstacle avoidance, parking, and real-robot experiments. The central claim is that projection-based constraints dramatically improve speed over generic constrained solvers and that ALSPG remains competitive with iLQR in the unconstrained case.","tokens_in":12002,"tokens_out":4931,"duration_ms":45804,"significance":"If validated, this is a practically useful contribution: it offers an open-source, memory-light solver that exploits geometric structure, with repeated-trial simulation benchmarks and real-world demonstrations. The efficiency gains over iLQR and IPOPT in constrained tasks are plausible, and the availability of code, data, and videos is a strength. However, the paper gives no convergence guarantee for the nonconvex and nonlinear extension that is central to the obstacle-avoidance and parking experiments, and one reported experiment contains an internal inconsistency that currently undermines the correctness claim for projection-based chance constraints.","major_comments":[{"comment":"The reported 'constraint violation percentage of around 80%, as expected' is the inverse of the intended chance constraint: for η=0.8, a correctly enforced constraint μ^T f(q)+Ψ^{-1}(η)||Σ^{1/2}f(q)||≤0 should be violated in at most 20% of samples. As written, this experiment suggests either the SOC projection does not enforce the intended constraint or the evaluation criterion is mislabeled; in either case, it does not support the claim that ALSPG correctly handles projection-based chance constraints. Please correct the evaluation or rerun the experiment and report the actual violation rate.","section":"Section V-A, Robust IK experiment"},{"comment":"The augmented Lagrangian gradient in Section IV-B uses the identity ∇∥g(x)−Π(g(x))∥² = ∇g(x)^T(g(x)−Π(g(x))), which is stated to follow from the derivative of convex Euclidean projections. For the nonconvex sets used in the experiments (e.g., outside-sphere constraints, polytopic avoidance), this identity is not generally valid, and the projection may be nondifferentiable or multi-valued. The paper explicitly concedes that the convergence results from [12] cover only convex functions and convex sets and that the nonconvex extension is empirical. Because the obstacle-avoidance and parking experiments rely on such nonconvex constraints, the algorithm's correctness for exactly the class of constraints that motivates 'projection-based optimization' is not established. Please either provide a stationarity or convergence analysis for the nonconvex case, or explicitly restrict the claims and add constraint-satisfaction metrics for these experiments.","section":"Section IV-B, Eq. (7) and derivative identity"},{"comment":"The central efficiency comparison for constrained motion planning is partly confounded: the 'ALSPG without Proj.' baseline uses the same augmented Lagrangian framework but with generic constraint gradients, while 'SLSQP with Proj.' uses a different subproblem solver and is noted to call C++ functions. The table reports convergence time, but the text acknowledges the comparison is 'not necessarily fair.' To support the claim that ALSPG with projections is faster than second-order methods in constrained problems, please provide a more controlled comparison, for example by implementing all baselines in the same language and reporting wall-clock time in a single environment.","section":"Section V-C, Tables IV and V"}],"minor_comments":[{"comment":"Tables III and IV have the identical caption 'Comparison of MPC with iLQR and ALSPG for planar push,' but Table IV is actually the obstacle-avoidance motion planning comparison; please correct the caption.","section":"Tables III and IV"},{"comment":"The multiplier update at line 5 is written with mixed subscripts; please clarify that λ_{C_i} is a vector of multipliers and that the update applies componentwise for each constraint.","section":"Algorithm 3, line 5"},{"comment":"The large standard deviation in function evaluations (6459.4±3756.8 without projections) suggests high variability; please report median and quartiles in addition to mean ± standard deviation.","section":"Section V-A, Talos IK"},{"comment":"The real-world experiments are single demonstrations; the '100% success rate' claim for the chess robot over three months would be stronger if the number of trials and the definition of a failed trial were reported.","section":"Sections V-D through V-F"},{"comment":"The notation F(x0,u)∈C_x and h(F(x0,u),u)=0 duplicates the state trajectory constraints; please clarify whether h includes the dynamics constraints or is a separate set of equality constraints.","section":"Section IV-C, Eq. (9)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a robotics journal and the self-citation pattern is appropriate. The main technical risk is the Robust IK experiment: if the 80% violation rate is not a simple typo, it indicates a load-bearing correctness failure in the projection-based chance-constraint handling. Please ask the authors to clarify and, if needed, rerun that experiment before proceeding."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper builds a first-order solver, ALSPG, by combining augmented Lagrangian with spectral projected gradient subproblems and geometric projections, then wraps it in a direct-shooting formulation for optimal control. That specific combination is new relative to the cited prior work, and it is a genuinely practical contribution: the benchmarks against iLQR and IPOPT on planar push, parking, and IK tasks show real speedups, and the authors ship code, data, and videos. The real-robot demos on Franka, P-Rob, and the 1:10 car are a plus.\n\nThe soft spots are real but mostly addressable. The biggest one is the Robust IK experiment in Section V-A. With η=0.8, a correctly enforced chance constraint should be violated about 20% of the time; the paper reports around 80% violation, which is exactly inverted. Either the SOC projection is not enforcing the intended chance constraint, or the evaluation is mislabeled. Either way, that experiment does not support the claim that ALSPG handles this projection-based constraint class correctly. This is load-bearing for the paper's central narrative, so it needs to be fixed before acceptance.\n\nSecond, the convergence theory is openly conceded to hold only for convex functions and sets (Section IV-B), while the experiments rely on nonconvex extensions justified by \"we found in practice.\" That is an honest limitation, and I don't think it sinks the paper, but the claims should be scaled to match. The paper already does a decent job of framing this as a practical method, so it mostly needs a clearer statement that the theoretical guarantees are still open.\n\nMinor: the real-robot demonstrations are single-run and lack statistical detail, and the chess robot result (100% accuracy over three months) is reported without error bars or trial counts. Those are less concerning given the simulation benchmarks.\n\nOverall, the efficiency claims are plausible and likely to generalize to similar problems, but the inverted IK result means the constraint-handling claims are not fully established as written. The paper deserves a serious referee, and I think it can be made solid with a focused revision. I would send it to review and ask for a corrected robust IK experiment and a tightened convergence statement.","headline":"A useful, incremental solver paper with strong empirical coverage, but the robust IK experiment reports an inverted violation rate that needs fixing before the constraint-handling claims can be trusted.","tokens_in":12463,"tokens_out":1378,"would_cite":true,"duration_ms":14887,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that ALSPG, a first-order augmented Lagrangian method with spectral projected gradient descent and geometric projections, solves constrained robot motion planning problems significantly faster in real time than…","keywords":["motion planning","constrained optimization","geometric projections","augmented Lagrangian","spectral projected gradient descent","inverse kinematics","model predictive control","collision avoidance"],"falsifier":"Run ALSPG on a moderately sized nonconvex trajectory-optimization problem, such as a 7-DOF arm making a choice between multiple separated obstacle-corridor paths, from many random initial guesses and record whether the constraint residual $V(x,\\lambda,\\rho)$ drops to the tolerance $\\epsilon_2 = 10^{-4}$; if the method frequently stalls at high residuals or the penalty updates oscillate indefinitely on problems where a second-order solver like IPOPT succeeds, the practical convergence claim fails.","tokens_in":11578,"feed_emoji":"🤖","tokens_out":6347,"duration_ms":48929,"temperature":0.7,"pith_summary":"This paper proposes ALSPG, a first-order optimization method for constrained robot motion planning. It replaces generic constraint gradients with fast geometric projections onto sets such as boxes, spheres, polytopes, and learned implicit shapes. The authors show that this projection-based approach solves inverse kinematics, motion planning, and model predictive control problems faster than second-order solvers like iLQR and IPOPT, and demonstrate it on real 6-axis and 7-axis arms and a 1:10 scale car. A sympathetic reader should care because real-time feedback loops in robotics need solvers that are fast and easy to implement, and the paper offers a concrete recipe.","feed_headline":"Projection-based solver beats iLQR and IPOPT in motion planning","feed_subtitle":"First-order optimizer using geometric projections runs in real-time and beats second-order baselines.","key_machinery":"The geometric projection operator $\\Pi_{\\mathcal{C}}(\\cdot)$, the Euclidean projection onto a constraint set $\\mathcal{C}$, is the central object, and it enters through an augmented Lagrangian function where each constraint $g_i(x)\\in\\mathcal{C}_i$ contributes a term $\\frac{\\rho_i}{2}\\lVert g_i(x)+\\lambda_i/\\rho_i - \\Pi_{\\mathcal{C}_i}(g_i(x)+\\lambda_i/\\rho_i)\\rVert^2$. Because the projection appears inside the penalty, the gradient of the projection is never needed, only the gradient of $g_i$ times the residual vector. The SPG subproblem solver supplies spectral stepsizes and a non-monotone line search, and a direct-shooting recursion computes products with $\\nabla_u F(x_0,u)^\\top$ without forming the full Jacobian matrix.","core_discovery":"The central discovery is that many robotic constraints — joint limits, stability regions, collision-avoidance sets, goal regions — can be written as geometric sets with analytical or learned projections, and that an augmented Lagrangian algorithm using spectral projected gradient descent (SPG) as its subproblem solver can exploit these projections to converge much faster than treating the constraints generically. The method, ALSPG, iteratively minimizes an augmented Lagrangian whose penalty terms measure the distance of constraint outputs to their target sets, and updates multipliers and penalty weights based on the projection residuals. In experiments, ALSPG with projections uses far fewer function and Jacobian evaluations than iLQR, IPOPT, and SLSQP baselines, and the paper reports real-time performance on a Franka arm, a P-Rob arm, and a 1:10 scale car. The paper also reports that in unconstrained problems ALSPG remains competitive with iLQR.","pith_inferences":["The measured speedups come from specific benchmark problems and hardware, so extending the same projection-based pattern to other structured constraints, such as contact manifolds or learned obstacle fields, would be a natural stress test of the approach's generality.","The paper leaves the convergence theory for nonconvex structured constraints open; finding a proof or a counterexample could either certify the method for safety-critical control or delimit where it should not be trusted.","Combining ALSPG with sample-based or learned dynamics could yield a hybrid planner that keeps the real-time constraint handling while inheriting model-free robustness, a direction the paper's conclusion hints at.","Because projections need not be convex, the same formulation might handle obstacle-exterior (nonconvex) constraints, which are common in navigation, without requiring convex decomposition."],"forward_implications":["Constrained inverse kinematics, obstacle-avoidance planning, and model predictive control can run inside the feedback loop on an onboard computer rather than purely offline.","Projection-based constraint handling reduces the cost of adding many geometric constraints, making richer task specifications such as boxes, half-spaces, cones, and collision sets practical in a single solver.","Because ALSPG needs only first-order derivatives and small memory, it can be embedded in resource-constrained robot controllers and onboard hardware.","The method provides a common solver for IK, planning, and MPC, potentially simplifying robot programming stacks that currently mix special-purpose solvers.","Competitiveness in the unconstrained case suggests the same solver can serve both constrained and unconstrained modes without swapping tools."],"supporting_citations":[{"why":"Supplies the SPG algorithm with spectral stepsizes and non-monotone line search that ALSPG uses as its subproblem solver.","marker":"[9]"},{"why":"Provides the augmented Lagrangian framework for structured geometric constraints that ALSPG extends to multiple nonlinear constraints.","marker":"[12]"},{"why":"Provides the direct-shooting recursion for matrix-vector products with the trajectory Jacobian and is the closest prior work using projected-gradient subproblems.","marker":"[19]"},{"why":"Supplies the Euclidean projection theory and the derivative identity used to compute the augmented Lagrangian gradient without projection gradients.","marker":"[15]"},{"why":"The iLQR baseline against which ALSPG's unconstrained and MPC performance is compared.","marker":"[6]"},{"why":"The IPOPT baseline solver used in parking and collision-avoidance benchmarks.","marker":"[4]"},{"why":"The SLSQP implementation from SciPy used as a second-order subproblem solver baseline in projection comparisons.","marker":"[22]"},{"why":"The OBCA algorithm with IPOPT used as a baseline in parking scenarios.","marker":"[23]"}],"fun_headline_variants":["ALSPG uses geometric projections for real-time motion planning","Projection-based optimizer beats iLQR and IPOPT in planning","Geometric projections make robot motion planning real-time","First-order method with projections rivals second-order solvers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's practical speedups rest on the assumption that the augmented Lagrangian iterations converge for general nonlinear and nonconvex robot constraints, which is not proven; the paper states that the theory in [12] covers convex cases and that \"we found in practice that the algorithm is powerful enough to extend to more general cases.\"","fun_headline_variants_meta":{"raw":{"variants":["ALSPG uses geometric projections for real-time motion planning","Projection-based optimizer beats iLQR and IPOPT in planning","Geometric projections make robot motion planning real-time","First-order method with projections rivals second-order solvers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1382,"prompt_tokens":922,"completion_tokens":460,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":406}},"tokens_in":538,"tokens_out":460,"duration_ms":5226,"temperature":1.0,"reasoning_tokens":406,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:09:25.848287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ALSPG on a moderately sized nonconvex trajectory-optimization problem, such as a 7-DOF arm making a choice between multiple separated obstacle-corridor paths, from many random initial guesses and record whether the constraint residual $V(x,\\lambda,\\rho)$ drops to the tolerance $\\epsilon_2 = 10^{-4}$; if the method frequently stalls at high residuals or the penalty updates oscillate indefinitely on problems where a second-order solver like IPOPT succeeds, the practical convergence claim fails.","supporting_citations":[{"cited_title":"Spectral projected gradient methods: review and perspectives,","cited_arxiv_id":null,"evidence_quote":"Supplies the SPG algorithm with spectral stepsizes and non-monotone line search that ALSPG uses as its subproblem solver."},{"cited_title":"An augmented la- grangian method for optimization problems with structured geometric constraints,","cited_arxiv_id":null,"evidence_quote":"Provides the augmented Lagrangian framework for structured geometric constraints that ALSPG extends to multiple nonlinear constraints."},{"cited_title":"A projected gradient and constraint linearization method for nonlinear model pre- dictive control,","cited_arxiv_id":null,"evidence_quote":"Provides the direct-shooting recursion for matrix-vector products with the trajectory Jacobian and is the closest prior work using projected-gradient subproblems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Euclidean projection theory and the derivative identity used to compute the augmented Lagrangian gradient without projection gradients."},{"cited_title":"Second-order differential dynamic programming for whole-body mpc of legged robots,","cited_arxiv_id":null,"evidence_quote":"The iLQR baseline against which ALSPG's unconstrained and MPC performance is compared."},{"cited_title":"An interior point algorithm for large-scale nonlinear opti- mization with applications in process engineering,","cited_arxiv_id":null,"evidence_quote":"The IPOPT baseline solver used in parking and collision-avoidance benchmarks."},{"cited_title":"Scipy 1.0: fundamental algorithms for scientific computing in python,","cited_arxiv_id":null,"evidence_quote":"The SLSQP implementation from SciPy used as a second-order subproblem solver baseline in projection comparisons."}],"review_version":1}